Study On Data Mining Technology's Application In Hydrological Forecast And Reservor Operation

As the vital step of knowledge discovery, data mining is the process of getting the unknown, valuable and operable information for decision-making support. There are a mass of hydroiogical, reservoir operation and forecasting data in the region of flood control in China. How to fully analyze and mine those data via various intelligent algorithms to formulate accordingly hydroiogical forecast and reservoir operation algorithm for precise forecast and rational operation is of crucial importance.Combining with the features of clustering, classification, association analysis, etc., this paper mainly discusses the application of those methods to hydroiogical forecast, reservoir operation and combined forecast. Meanwhile, for a more effective data mining analysis, this paper demonstrates the structure and application of hydroiogical data warehouse, which is integrated with various data mining applications to formulate a hydroiogical data mining system. The Contents and results of this paper are listed, as follows.(1) Aiming at the characteristics of data mining-based forecast and operation, integrating with the condition of flood mitigation in China, the structural model of hydroiogical data warehouse, which is based on forecast and operation, is established in this paper, and then the structure, function, data storage model and realization are studied to manage and analyze the massive hydroiogical data, which offers data support and promotes data-mining efficiency for data mining-based forecast and operation system. Then, the hydroiogical data mining system, which includes data layer, organization layer, mining layer and decision layer, is established. Different layer has its own functions for data pre-process, date mining to knowledge expression under different stages of hydroiogical data mining to formulate a whole system.(2) Taking the river flood propagation forecast as an example, the data mining-based hydroiogical model is studied. The river flood data is pre-processed by using hydroiogical data warehouse, meanwhile a T-S fuzzy inference model-based model is proposed for inferring the downstream flow based on the conditions of upstream branches. Aiming at the "curse of dimensionality" problems with the increase of inferring conditions, association analysis is used to determine the combination of upstream historical data with frequent occurrence, i.e., inferring rules, to combine or delete the low flow combinations or rules with scarce occurrence, then the optimizing model is used to analyze the rule set and specify the parameters, and finally the flood propagation model based on fuzzy inference and association analysis is formulated, whichpromotes the precision of high flow forecast.(3) Due to different hydrological forecast model has different feature, there exists no forecast model. which is suitable for all circumstances. Aiming at this problem, a combined forecast model, which is based on the fuzzy optimizing model proposed by Professor Chen Shouyu, is presented in this paper. Taking the relative errors of flood peak, flood peak time and How process of different model as indices for fuzzy optimization, the fuzzy pattern recognition analysis is carried out towards historical forecasted data to get the suitability of various models under different circumstances. During real-time forecast, the combined forecast is carried out according to the suitability of various models, and the more reasonable forecasting results are obtained compared with those by single forecasting model.(4) Based on the research focuses in the research field of reservoir operation, a reservoir operation model, i.e., decision tree model, which is based on the data-mining for historical hydrological and operation data, is proposed. As the result of data-mining, this model reflects the rules hidden in the mass of historical hydrological and operation data. At the same time, if the historical data and expert experience are qualified, the hydrological forecast result, weather forecast result and ecological factors can be considered in the rules, which will make the decision-making more reasonable. Compared with the extensively used operation graph at present, operation tree is more reasonable with extensive considerations and direct expression for rules, and is the effective approach for reasonably using operation experience.(5) Compared with the combined forecast method discussed in Chapter 4, the Bayesian theory-based combined forecast method needs no more historical forecast data for training the combined model. However, the expert experience is used to specify the distribution type of the model, then through Bayesian analysis, the posterior distribution of forecast results are turned into the likelihoods of the combined forecast result, then using the Markov Chain Monte Carlo (MCMC) methodologies to sampling simulate the combined forecast result. In the process of sampling, the real-time correction method based time-serial are added to the Gibbs sampling of MCMC to get the better combined forecast result.Finally, the conclusion is made, and the problems for further study are reviewed.